def test_squeeze(is_func, axis): a = np.arange(10)[None, :, None, None] d = da.from_array(a, chunks=(1, 3, 1, 1)) if is_func: a_s = np.squeeze(a, axis=axis) d_s = da.squeeze(d, axis=axis) else: a_s = a.squeeze(axis=axis) d_s = d.squeeze(axis=axis) assert_eq(d_s, a_s) assert same_keys(d_s, da.squeeze(d, axis=axis)) if axis is None: axis = tuple(range(a.ndim)) else: axis = axis if isinstance(axis, tuple) else (axis,) axis = tuple(i % a.ndim for i in axis) axis = tuple( i for i, c in enumerate(d.chunks) if i in axis and len(c) == 1 ) exp_d_s_chunks = tuple( c for i, c in enumerate(d.chunks) if i not in axis ) assert d_s.chunks == exp_d_s_chunks
def tifffile_dask_backend(image_filepath, largest_series, preprocessing, force_rgb=None): """ Read image with tifffile and use dask to read data into memory Parameters ---------- image_filepath: str path to the image file largest_series: int index of the largest series in the image preprocessing: whether to do some read-time pre-processing - greyscale conversion (at the tile level) - read individual or range of channels (at the tile level) Returns ------- image: sitk.Image image ready for other registration pre-processing """ print("using dask backend") zarr_series = imread(image_filepath, aszarr=True, series=largest_series) zarr_store = zarr.open(zarr_series) dask_im = da.squeeze(da.from_zarr(zarr_get_base_pyr_layer(zarr_store))) if force_rgb is None: is_rgb = guess_rgb(dask_im.shape) is_interleaved = True if is_rgb else False elif force_rgb is True and guess_rgb(dask_im.shape) is False: is_rgb = True is_interleaved = False if is_rgb: if preprocessing is not None: image = grayscale(dask_im, is_interleaved=is_interleaved).compute() image = sitk.GetImageFromArray(image) else: image = dask_im.compute() if is_interleaved is False: image = np.rollaxis(image, 0, 3) image = sitk.GetImageFromArray(image, isVector=True) elif len(dask_im.shape) == 2: image = sitk.GetImageFromArray(dask_im.compute()) else: if preprocessing is not None: if (preprocessing.get("ch_indices") is not None and len(dask_im.shape) > 2): chs = np.asarray(preprocessing.get('ch_indices')) dask_im = dask_im[chs, :, :] image = sitk.GetImageFromArray(np.squeeze(dask_im.compute())) return image
def tifffile_dask_backend(image_filepath, largest_series, preprocessing, force_rgb=None): """ Read image with tifffile and use dask to read data into memory Parameters ---------- image_filepath: str path to the image file largest_series: int index of the largest series in the image preprocessing: whether to do some read-time pre-processing - greyscale conversion (at the tile level) - read individual or range of channels (at the tile level) Returns ------- image: sitk.Image image ready for other registration pre-processing """ print("using dask backend") zarr_series = imread(image_filepath, aszarr=True, series=largest_series) zarr_store = zarr.open(zarr_series) dask_im = da.squeeze(da.from_zarr(zarr_get_base_pyr_layer(zarr_store))) return read_preprocess_array(dask_im, preprocessing=preprocessing, force_rgb=force_rgb)
def _matvec(self, x): x = da.squeeze(x.reshape(self.nsl, self.ny, self.nz)) if self.chunks[0] is not None: x = x.rechunk(self.chunks[0]) if self.nz == 1: x = x[..., np.newaxis] y = da.matmul(self.G, x) return y.ravel()
def tf_zarr_read_single_ch(image_filepath, channel_idx, is_rgb, is_rgb_interleaved=False): """ Reads a single channel using zarr or dask in combination with tifffile Parameters ---------- image_filepath:str file path to image channel_idx:int index of the channel to be read is_rgb:bool whether image is rgb interleaved Returns ------- im:np.ndarray image as a np.ndarray """ largest_series = tf_get_largest_series(image_filepath) zarr_im = zarr.open( imread(image_filepath, aszarr=True, series=largest_series)) zarr_im = zarr_get_base_pyr_layer(zarr_im) try: im = da.squeeze(da.from_zarr(zarr_im)) if is_rgb and is_rgb_interleaved is True: im = im[:, :, channel_idx].compute() elif len(im.shape) > 2: im = im[channel_idx, :, :].compute() else: im = im.compute() except ValueError: im = zarr_im if is_rgb is True and is_rgb_interleaved is True: im = im[:, :, channel_idx] elif len(im.shape) > 2: im = im[channel_idx, :, :].compute() else: im = im.compute() return im
def apply(X: Array, YP: Array, BX: Array, BYP: Array) -> Array: # Collapse selected variant blocks and alphas into single # new covariate dimension assert YP.shape[2] == BYP.shape[2] n_group_covar = n_covar + BYP.shape[2] * n_alpha_1 BYP = BYP.reshape((n_outcome, n_sample_block, -1)) BG = da.concatenate((BX, BYP), axis=-1) BG = BG.rechunk((-1, None, -1)) assert_block_shape(BG, 1, n_sample_block, 1) assert_chunk_shape(BG, n_outcome, 1, n_group_covar) assert_array_shape(BG, n_outcome, n_sample_block, n_group_covar) YP = YP.reshape((n_outcome, n_sample, -1)) XYP = da.broadcast_to(X, (n_outcome, n_sample, n_covar)) XG = da.concatenate((XYP, YP), axis=-1) XG = XG.rechunk((-1, None, -1)) assert_block_shape(XG, 1, n_sample_block, 1) assert_chunk_shape(XG, n_outcome, sample_chunks[0], n_group_covar) assert_array_shape(XG, n_outcome, n_sample, n_group_covar) YG = da.map_blocks( # Block chunks: # (n_outcome, sample_chunks[0], n_group_covar) @ # (n_outcome, n_group_covar, 1) [after transpose] lambda x, b: x @ b.transpose((0, 2, 1)), XG, BG, chunks=(n_outcome, sample_chunks, 1), ) assert_block_shape(YG, 1, n_sample_block, 1) assert_chunk_shape(YG, n_outcome, sample_chunks[0], 1) assert_array_shape(YG, n_outcome, n_sample, 1) YG = da.squeeze(YG, axis=-1).T assert_block_shape(YG, n_sample_block, 1) assert_chunk_shape(YG, sample_chunks[0], n_outcome) assert_array_shape(YG, n_sample, n_outcome) return YG
def _resample_tile( self, ch_idx: int, tile_resampler: sitk.ResampleImageFilter, x_max: int, x_min: int, y_max: int, y_min: int, ) -> Optional[sitk.Image]: """Resample tile or don't if it is outside of the moving image space.""" if x_min == 0 and x_max == 0: return if y_min == 0 and y_max == 0: return if self.reg_image.is_rgb: image = self.reg_image.dask_image[y_min:y_max, x_min:x_max, :] image = sitk.GetImageFromArray(image, isVector=True) elif self.reg_image.n_ch == 1: image = da.squeeze(self.reg_image.dask_image)[ y_min:y_max, x_min:x_max ] image = sitk.GetImageFromArray(image, isVector=False) else: image = self.reg_image.dask_image[ch_idx, y_min:y_max, x_min:x_max] image = sitk.GetImageFromArray(image, isVector=False) image.SetSpacing((self.reg_image.image_res, self.reg_image.image_res)) image.SetOrigin( image.TransformIndexToPhysicalPoint([int(x_min), int(y_min)]) ) tile_resampled = tile_resampler.Execute(image) return tile_resampled
def _center_x(self, x, dx, transpose: bool = False) -> da.core.Array: """ Centers the product of matrix multiplication instead of center the matrix Let A be a matrix of shape (n by p) with non zero column means, U of shape (p,). Matrix B could be constructed as follows with zero column mean. B = A - 1'U where 1 is a 1 vector. And 1'U is an outer product of shape (n by p) However, this is inefficient if only the matrix product of B, with a matrix x is needed. Instead `_center_x` implements: Ax - Ux ^ ^- dx being passed in, | x being passed in with efficient broadcasting. Parameters ---------- x : array_like Usually the product of Ax that needs to be center dx : array_like Usually the original x before being multiplied by A transpose : bool Flag whether to indicate if A'x or Ax. Adjusts dimensions Returns ------- x_centered: array_like """ if transpose: # Computes mu1'x_k_h return x - da.squeeze( da.outer(self._array_moment.center_vector, dx.sum(axis=0))) else: return x - self._array_moment.center_vector.dot(dx)
def to_raster(data, filename, readxsize=None, readysize=None, use_dask_store=False, separate=False, out_block_type='zarr', keep_blocks=False, verbose=0, overwrite=False, gdal_cache=512, scheduler='mpool', n_jobs=1, n_workers=None, n_threads=None, n_chunks=None, overviews=False, resampling='nearest', use_client=False, address=None, total_memory=48, **kwargs): """ Writes a ``dask`` array to a raster file Args: data (DataArray): The ``xarray.DataArray`` to write. filename (str): The output file name to write to. readxsize (Optional[int]): The size of column chunks to read. If not given, ``readxsize`` defaults to Dask chunk size. readysize (Optional[int]): The size of row chunks to read. If not given, ``readysize`` defaults to Dask chunk size. separate (Optional[bool]): Whether to write blocks as separate files. Otherwise, write to a single file. use_dask_store (Optional[bool]): Whether to use ``dask.array.store`` to save with Dask task graphs. out_block_type (Optional[str]): The output block type. Choices are ['gtiff', 'zarr']. Only used if ``separate`` = ``True``. keep_blocks (Optional[bool]): Whether to keep the blocks stored on disk. Only used if ``separate`` = ``True``. verbose (Optional[int]): The verbosity level. overwrite (Optional[bool]): Whether to overwrite an existing file. gdal_cache (Optional[int]): The ``GDAL`` cache size (in MB). scheduler (Optional[str]): The ``concurrent.futures`` scheduler to use. Choices are ['processes', 'threads', 'mpool']. mpool: process pool of workers using ``multiprocessing.Pool`` processes: process pool of workers using ``concurrent.futures`` threads: thread pool of workers using ``concurrent.futures`` n_jobs (Optional[int]): The total number of parallel jobs. n_workers (Optional[int]): The number of processes. n_threads (Optional[int]): The number of threads. n_chunks (Optional[int]): The chunk size of windows. If not given, equal to ``n_workers`` x 50. overviews (Optional[bool or list]): Whether to build overview layers. resampling (Optional[str]): The resampling method for overviews when ``overviews`` is ``True`` or a ``list``. Choices are ['average', 'bilinear', 'cubic', 'cubic_spline', 'gauss', 'lanczos', 'max', 'med', 'min', 'mode', 'nearest']. use_client (Optional[bool]): Whether to use a ``dask`` client. address (Optional[str]): A cluster address to pass to client. Only used when ``use_client`` = ``True``. total_memory (Optional[int]): The total memory (in GB) required when ``use_client`` = ``True``. kwargs (Optional[dict]): Additional keyword arguments to pass to ``rasterio.write``. Returns: ``dask.delayed`` object Examples: >>> import geowombat as gw >>> >>> # Use 8 parallel workers >>> with gw.open('input.tif') as ds: >>> gw.to_raster(ds, 'output.tif', n_jobs=8) >>> >>> # Use 4 process workers and 2 thread workers >>> with gw.open('input.tif') as ds: >>> gw.to_raster(ds, 'output.tif', n_workers=4, n_threads=2) >>> >>> # Control the window chunks passed to concurrent.futures >>> with gw.open('input.tif') as ds: >>> gw.to_raster(ds, 'output.tif', n_workers=4, n_threads=2, n_chunks=16) >>> >>> # Compress the output and build overviews >>> with gw.open('input.tif') as ds: >>> gw.to_raster(ds, 'output.tif', n_jobs=8, overviews=True, compress='lzw') """ if MKL_LIB: __ = MKL_LIB.MKL_Set_Num_Threads(n_threads) pfile = Path(filename) if scheduler.lower() == 'mpool': pool_executor = multi.Pool else: pool_executor = concurrent.futures.ProcessPoolExecutor if scheduler.lower( ) == 'processes' else concurrent.futures.ThreadPoolExecutor if overwrite: if pfile.is_file(): pfile.unlink() if pfile.is_file(): logger.warning(' The output file already exists.') return if not is_dask_collection(data.data): logger.exception(' The data should be a dask array.') if use_client: if address: cluster_object = _cluster_dummy else: cluster_object = LocalCluster client_object = Client else: cluster_object = _cluster_dummy client_object = _client_dummy if isinstance(n_workers, int) and isinstance(n_threads, int): n_jobs = n_workers * n_threads else: n_workers = n_jobs n_threads = 1 mem_per_core = int(total_memory / n_workers) if not isinstance(n_chunks, int): n_chunks = n_workers * 50 if not isinstance(readxsize, int): readxsize = data.gw.col_chunks if not isinstance(readysize, int): readysize = data.gw.row_chunks chunksize = (data.gw.row_chunks, data.gw.col_chunks) # Force tiled outputs with no file sharing kwargs['sharing'] = False if data.gw.tiled: kwargs['tiled'] = True if 'compress' in kwargs: # Store the compression type because # it is removed in concurrent writing compress = True compress_type = kwargs['compress'] del kwargs['compress'] elif isinstance(data.gw.compress, str) and data.gw.compress.lower() in ['lzw', 'deflate']: compress = True compress_type = data.gw.compress else: compress = False if 'nodata' not in kwargs: if isinstance(data.gw.nodata, int) or isinstance( data.gw.nodata, float): kwargs['nodata'] = data.gw.nodata if 'blockxsize' not in kwargs: kwargs['blockxsize'] = data.gw.col_chunks if 'blockysize' not in kwargs: kwargs['blockysize'] = data.gw.row_chunks if 'bigtiff' not in kwargs: kwargs['bigtiff'] = data.gw.bigtiff if 'driver' not in kwargs: kwargs['driver'] = data.gw.driver if 'count' not in kwargs: kwargs['count'] = data.gw.nbands if 'width' not in kwargs: kwargs['width'] = data.gw.ncols if 'height' not in kwargs: kwargs['height'] = data.gw.nrows if separate: d_name = pfile.parent sub_dir = d_name.joinpath('sub_tmp_') zarr_file = sub_dir.joinpath('data.zarr').as_posix() sub_dir.mkdir(parents=True, exist_ok=True) root = zarr.open(zarr_file, mode='w') else: root = None if verbose > 0: logger.info(' Creating the file ...\n') with rio.open(filename, mode='w', **kwargs) as rio_dst: pass if verbose > 0: logger.info(' Writing data to file ...\n') with rio.Env(GDAL_CACHEMAX=gdal_cache): if not use_dask_store: windows = get_window_offsets(data.gw.nrows, data.gw.ncols, readysize, readxsize, return_as='list') n_windows = len(windows) # Iterate over the windows in chunks for wchunk in range(0, n_windows, n_chunks): window_slice = windows[wchunk:wchunk + n_chunks] n_windows_slice = len(window_slice) if verbose > 0: logger.info(' Windows {:,d}--{:,d} of {:,d} ...'.format( wchunk + 1, wchunk + n_windows_slice, n_windows)) if len(data.shape) == 2: data_gen = ((data[w.row_off:w.row_off + w.height, w.col_off:w.col_off + w.width], filename, w, n_threads, separate, chunksize, root) for w in window_slice) elif len(data.shape) == 3: data_gen = ((data[:, w.row_off:w.row_off + w.height, w.col_off:w.col_off + w.width], filename, w, n_threads, separate, chunksize, root) for w in window_slice) else: data_gen = ((data[:, :, w.row_off:w.row_off + w.height, w.col_off:w.col_off + w.width], filename, w, n_threads, separate, chunksize, root) for w in window_slice) with pool_executor(n_workers) as executor: if scheduler == 'mpool': for zarr_file in tqdm(executor.imap_unordered( _write_xarray, data_gen), total=n_windows_slice): pass else: for zarr_file in tqdm(executor.map( _write_xarray, data_gen), total=n_windows_slice): pass # if overviews: # # if not isinstance(overviews, list): # overviews = [2, 4, 8, 16] # # if resampling not in ['average', 'bilinear', 'cubic', 'cubic_spline', # 'gauss', 'lanczos', 'max', 'med', 'min', 'mode', 'nearest']: # # logger.warning(" The resampling method is not supported by rasterio. Setting to 'nearest'") # # resampling = 'nearest' # # if verbose > 0: # logger.info(' Building pyramid overviews ...') # # rio_dst.build_overviews(overviews, getattr(Resampling, resampling)) # rio_dst.update_tags(ns='overviews', resampling=resampling) else: with cluster_object( n_workers=n_workers, threads_per_worker=n_threads, scheduler_port=0, processes=False, memory_limit='{:d}GB'.format(mem_per_core)) as cluster: cluster_address = address if address else cluster with client_object(address=cluster_address) as client: with WriteDaskArray(filename, overwrite=overwrite, separate=separate, out_block_type=out_block_type, keep_blocks=keep_blocks, gdal_cache=gdal_cache, **kwargs) as dst: # Store the data and return a lazy evaluator res = da.store(da.squeeze(data.data), dst, lock=False, compute=False) if verbose > 0: logger.info(' Writing data to file ...') # Send the data to file # # *Note that the progress bar will # not work with a client. if use_client: res.compute(num_workers=n_jobs) else: with ProgressBar(): res.compute(num_workers=n_jobs) if verbose > 0: logger.info(' Finished writing data to file.') out_block_type = dst.out_block_type keep_blocks = dst.keep_blocks zarr_file = dst.zarr_file sub_dir = dst.sub_dir if compress: if verbose > 0: logger.info(' Compressing output file ...') if separate: group_keys = list(root.group_keys()) n_groups = len(group_keys) if out_block_type.lower() == 'zarr': # root = zarr.open(zarr_file, mode='r') open_file = zarr_file else: outfiles = sorted( fnmatch.filter(os.listdir(sub_dir), '*.tif')) outfiles = [os.path.join(sub_dir, fn) for fn in outfiles] # data_gen = ((fn, None, 'gtiff') for fn in outfiles) kwargs['compress'] = compress_type n_windows = len(group_keys) # Compress into one file with rio.open(filename, mode='w', **kwargs) as dst_: # Iterate over the windows in chunks for wchunk in range(0, n_groups, n_chunks): group_keys_slice = group_keys[wchunk:wchunk + n_chunks] n_windows_slice = len(group_keys_slice) if verbose > 0: logger.info( ' Windows {:,d}--{:,d} of {:,d} ...'.format( wchunk + 1, wchunk + n_windows_slice, n_windows)) ################################################ data_gen = ((open_file, group, 'zarr') for group in group_keys_slice) # for f in tqdm(executor.map(_compressor, data_gen), total=n_windows_slice): # pass # # futures = [executor.submit(_compress_dummy, iter_[0], iter_[1], None) for iter_ in data_gen] # # for f in tqdm(concurrent.futures.as_completed(futures), total=n_windows_slice): # # out_window, out_block = f.result() # # dst_.write(np.squeeze(out_block), # window=out_window, # indexes=out_indexes_) ################################################ # data_gen = ((root, group, 'zarr') for group in group_keys_slice) # for f, g, t in tqdm(data_gen, total=n_windows_slice): # # out_window, out_indexes, out_block = _block_read_func(f, g, t) # executor.map(_block_write_func, data_gen) with concurrent.futures.ProcessPoolExecutor( max_workers=n_workers) as executor: # Submit all of the tasks as futures futures = [ executor.submit(_block_read_func, f, g, t) for f, g, t in data_gen ] for f in tqdm( concurrent.futures.as_completed(futures), total=n_windows_slice): out_window, out_indexes, out_block = f.result() dst_.write(out_block, window=out_window, indexes=out_indexes) futures = None if not keep_blocks: shutil.rmtree(sub_dir) else: p = Path(filename) d_name = p.parent f_base, f_ext = os.path.splitext(p.name) ld = string.ascii_letters + string.digits rstr = ''.join(random.choice(ld) for i in range(0, 9)) temp_file = d_name.joinpath('{}_temp_{}{}'.format( f_base, rstr, f_ext)) compress_raster(filename, temp_file.as_posix(), n_jobs=n_jobs, gdal_cache=gdal_cache, compress=compress_type) temp_file.rename(filename) if verbose > 0: logger.info(' Finished compressing') if verbose > 0: logger.info('\nFinished writing the data.')
def as_array(self, axes=None, stitched=False, verbose=True, **kwargs): """ Read all data image data as one big Dask array with last two axes as y, x and preceeding axes depending on data. The dask array is made up of memory-mapped numpy arrays, so the dataset does not need to be able to fit into RAM. If the data doesn't fully fill out the array (e.g. not every z-slice collected at every time point), zeros will be added automatically. To convert data into a numpy array, call np.asarray() on the returned result. However, doing so will bring the data into RAM, so it may be better to do this on only a slice of the array at a time. Parameters ---------- axes : list list of axes names over which to iterate and merge into a stacked array. If None, all axes will be used stitched : bool If true and tiles were acquired in a grid, lay out adjacent tiles next to one another (Default value = False) verbose : bool If True print updates on progress loading the image **kwargs : names and integer positions of axes on which to slice data Returns ------- dataset : dask array """ w = self.image_height if not stitched else self._tile_width h = self.image_height if not stitched else self._tile_height self._empty_tile = (np.zeros( (h, w), self.dtype) if self.bytes_per_pixel != 3 else np.zeros( (h, w, 3), self.dtype)) self._count = 1 total = np.prod([len(v) for v in self.axes.values()]) def recurse_axes(loop_axes, point_axes): """ Used to create a nested list of images, with each nesting level corresponding to a particular axis. Each time this function is recursively called, it will descend one level deeper. The recursive calls can be thought of as a tree structure, where each depth level of the tree is one axis, and it has a branch (i.e. a subsequent call of recurse_axes) corresponding to every value of the the next axis. :param loop_axes: The remaining axes that need to be looped over (i.e. the innermost ones) :param point_axes: The axes that have been assigned values already by a previous call of this function :return: Nested list of images """ if len(loop_axes.values()) == 0: # There are no more axes over which to loop (i.e. we're at the maximum depth), so return # the image defined by point_axes, or a blank image if it is undefined (so that the full # nested list will have the expected rectangular shape) if verbose: print("\rAdding data chunk {} of {}".format( self._count, total), end="") self._count += 1 if None not in point_axes.values() and self.has_image( **point_axes): recurse_axes.empty = False # track that actual data was read if stitched: img = self.read_image(**point_axes, memmapped=True) if self.half_overlap[0] != 0: img = img[ self.half_overlap[0]:-self.half_overlap[0], self.half_overlap[1]:-self.half_overlap[1], ] return img else: return self.read_image(**point_axes, memmapped=True) else: # return np.zeros((self.image_height, self.image_width), self.dtype) return self._empty_tile else: # Still have axes over which to loop # do row and col first because it makes stitching faster if "row" in loop_axes.keys() and stitched: axis = "row" elif "column" in loop_axes.keys() and stitched: axis = "column" else: # Take the next axis in the list that needs to be looped over axis = list(loop_axes.keys())[0] # copy so multiple calls dont collide on the same data structure remaining_loop_axes = loop_axes.copy() if (axis == "row" or axis == "column") and stitched: # do these both at once del remaining_loop_axes["row"] del remaining_loop_axes["column"] else: # remove because this axis is now being assigned a point value del remaining_loop_axes[axis] if (axis == "row" or axis == "column") and stitched: # Do stitching along existing axis # Stitch tiles acquired in a grid (i.e. data acquired by Micro-Magellan or in multi-res mode) self.half_overlap = (self.overlap[0] // 2, self.overlap[1] // 2) # get spatial layout of position indices row_values = np.array(list(self.axes["row"])) column_values = np.array(list(self.axes["column"])) # make nested list of rows and columns blocks = [] for row in row_values: blocks.append([]) for column in column_values: valed_axes = point_axes.copy() if verbose: print( "\rAdding data chunk {} of {}".format( self._count, total), end="", ) valed_axes["row"] = row valed_axes["column"] = column blocks[-1].append( da.stack( recurse_axes(remaining_loop_axes, valed_axes))) rgb = self.bytes_per_pixel == 3 and self.dtype == np.uint8 if rgb: stitched_array = np.concatenate( [ np.concatenate( row, axis=len(blocks[0][0].shape) - 2) for row in blocks ], axis=len(blocks[0][0].shape) - 3, ) else: stitched_array = da.block(blocks) return stitched_array else: # Do stacking along new axis (i.e. not stiching along exisitng) blocks = [] # Loop through every value of the next axis (i.e. create new branches of the tree) for val in loop_axes[axis]: # Copy to avoid unexpected errors by multiple calls valed_axes = point_axes.copy() # Move this axis from one that needs to be looped over to one that has a discrete value. valed_axes[axis] = val blocks.append( recurse_axes(remaining_loop_axes, valed_axes)) return blocks if axes is None: axes = self.axes.keys() axes_to_slice = kwargs axes_to_stack_or_stitch = { key: self.axes[key] for key in axes if key not in kwargs.keys() } recurse_axes.empty = True blocks = recurse_axes(axes_to_stack_or_stitch, axes_to_slice) if recurse_axes.empty: # No actual data in any of the tiles return None if verbose: print( "\rStacking tiles... " ) # extra space otherwise there is no space after the "Adding data chunk {} {}" # import time # s = time.time() array = da.stack(blocks, allow_unknown_chunksizes=False) # e = time.time() # print(e - s) if verbose: print("\rDask array opened") # remove singleton axes array = da.squeeze(array) return array
def make_regression(n_samples=100, n_features=100, n_informative=10, n_targets=1, bias=0.0, effective_rank=None, tail_strength=0.5, noise=0.0, shuffle=False, coef=False, random_state=None, n_parts=1, n_samples_per_part=None): """Generate a random regression problem. The input set can either be well conditioned (by default) or have a low rank-fat tail singular profile. The output is generated by applying a (potentially biased) random linear regression model with "n_informative" nonzero regressors to the previously generated input and some gaussian centered noise with some adjustable scale. Parameters ---------- n_samples : int, optional (default=100) The number of samples. n_features : int, optional (default=100) The number of features. n_informative : int, optional (default=10) The number of informative features, i.e., the number of features used to build the linear model used to generate the output. n_targets : int, optional (default=1) The number of regression targets, i.e., the dimension of the y output vector associated with a sample. By default, the output is a scalar. bias : float, optional (default=0.0) The bias term in the underlying linear model. effective_rank : int or None, optional (default=None) if not None: The approximate number of singular vectors required to explain most of the input data by linear combinations. Using this kind of singular spectrum in the input allows the generator to reproduce the correlations often observed in practice. if None: The input set is well conditioned, centered and gaussian with unit variance. tail_strength : float between 0.0 and 1.0, optional (default=0.5) The relative importance of the fat noisy tail of the singular values profile if "effective_rank" is not None. noise : float, optional (default=0.0) The standard deviation of the gaussian noise applied to the output. shuffle : boolean, optional (default=False) Shuffle the samples and the features. coef : boolean, optional (default=False) If True, the coefficients of the underlying linear model are returned. random_state : int, CuPy RandomState instance, Dask RandomState instance or None (default) Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls. n_parts : int, optional (default=1) The number of parts of work. Returns ------- X : Dask-CuPy array of shape [n_samples, n_features] The input samples. y : Dask-CuPy array of shape [n_samples] or [n_samples, n_targets] The output values. coef : Dask-CuPy array of shape [n_features] or [n_features, n_targets], optional The coefficient of the underlying linear model. It is returned only if coef is True. """ n_informative = min(n_features, n_informative) rs = create_rs_generator(random_state) if n_samples_per_part is None: n_samples_per_part = max(1, int(n_samples / n_parts)) if effective_rank is None: # Randomly generate a well conditioned input set X = rs.standard_normal( (n_samples, n_features), chunks=(n_samples_per_part, (n_informative, n_features - n_informative))) else: # Randomly generate a low rank, fat tail input set X = make_low_rank_matrix(n_samples=n_samples, n_features=n_features, effective_rank=effective_rank, tail_strength=tail_strength, random_state=rs, n_parts=n_parts) X = X.rechunk({ 0: n_samples_per_part, 1: (n_informative, n_features - n_informative) }) # Generate a ground truth model with only n_informative features being non # zeros (the other features are not correlated to y and should be ignored # by a sparsifying regularizers such as L1 or elastic net) ground_truth = 100.0 * rs.standard_normal( (n_informative, n_targets), chunks=(n_samples_per_part, -1)) y = da.dot(X[:, :n_informative], ground_truth) + bias X = X.rechunk((None, -1)) if n_informative != n_features: zeroes = 0.0 * rs.standard_normal( (n_features - n_informative, n_targets)) ground_truth = da.concatenate([ground_truth, zeroes], axis=0) ground_truth = ground_truth.rechunk(-1) # Add noise if noise > 0.0: y += rs.normal(scale=noise, size=y.shape) # Randomly permute samples and features if shuffle: samples_indices = np.random.permutation(n_samples) X = X[samples_indices, :] y = y[samples_indices, :] features_indices = np.random.permutation(n_features) X = X[:, features_indices] ground_truth = ground_truth[features_indices, :] y = da.squeeze(y) if coef: ground_truth = da.squeeze(ground_truth) return X, y, ground_truth else: return X, y
def make_regression(n_samples=100, n_features=100, n_informative=10, n_targets=1, bias=0.0, effective_rank=None, tail_strength=0.5, noise=0.0, shuffle=False, coef=False, random_state=None, n_parts=1, n_samples_per_part=None, order='F', dtype='float32', client=None, use_full_low_rank=True): """ Generate a random regression problem. The input set can either be well conditioned (by default) or have a low rank-fat tail singular profile. The output is generated by applying a (potentially biased) random linear regression model with "n_informative" nonzero regressors to the previously generated input and some gaussian centered noise with some adjustable scale. Parameters ---------- n_samples : int, optional (default=100) The number of samples. n_features : int, optional (default=100) The number of features. n_informative : int, optional (default=10) The number of informative features, i.e., the number of features used to build the linear model used to generate the output. n_targets : int, optional (default=1) The number of regression targets, i.e., the dimension of the y output vector associated with a sample. By default, the output is a scalar. bias : float, optional (default=0.0) The bias term in the underlying linear model. effective_rank : int or None, optional (default=None) if not None: The approximate number of singular vectors required to explain most of the input data by linear combinations. Using this kind of singular spectrum in the input allows the generator to reproduce the correlations often observed in practice. if None: The input set is well conditioned, centered and gaussian with unit variance. tail_strength : float between 0.0 and 1.0, optional (default=0.5) The relative importance of the fat noisy tail of the singular values profile if "effective_rank" is not None. noise : float, optional (default=0.0) The standard deviation of the gaussian noise applied to the output. shuffle : boolean, optional (default=False) Shuffle the samples and the features. coef : boolean, optional (default=False) If True, the coefficients of the underlying linear model are returned. random_state : int, CuPy RandomState instance, Dask RandomState instance \ or None (default) Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls. n_parts : int, optional (default=1) The number of parts of work. order : str, optional (default='F') Row-major or Col-major dtype: str, optional (default='float32') dtype of generated data use_full_low_rank : boolean (default=True) Whether to use the entire dataset to generate the low rank matrix. If False, it creates a low rank covariance and uses the corresponding covariance to generate a multivariate normal distribution on the remaining chunks Returns ------- X : Dask-CuPy array of shape [n_samples, n_features] The input samples. y : Dask-CuPy array of shape [n_samples] or [n_samples, n_targets] The output values. coef : Dask-CuPy array of shape [n_features] \ or [n_features, n_targets], optional The coefficient of the underlying linear model. It is returned only if coef is True. Notes ----- Known Performance Limitations: 1. When `effective_rank` is set and `use_full_low_rank` is True, \ we cannot generate order `F` by construction, and an explicit \ transpose is performed on each part. This may cause memory to spike \ (other parameters make order `F` by construction) 2. When `n_targets > 1` and `order = 'F'` as above, we have to \ explicity transpose the `y` array. If `coef = True`, then we also \ explicity transpose the `ground_truth` array 3. When `shuffle = True` and `order = F`, there are memory spikes to \ shuffle the `F` order arrays .. note:: If out-of-memory errors are encountered in any of the above configurations, try increasing the `n_parts` parameter. """ client = get_client(client=client) n_informative = min(n_features, n_informative) rs = _create_rs_generator(random_state) if n_samples_per_part is None: n_samples_per_part = max(1, int(n_samples / n_parts)) data_chunksizes = [n_samples_per_part] * n_parts data_chunksizes[-1] += (n_samples % n_parts) data_chunksizes = tuple(data_chunksizes) if effective_rank is None: # Randomly generate a well conditioned input set if order == 'F': X = _f_order_standard_normal(client, rs, data_chunksizes, n_features, dtype) elif order == 'C': X = rs.standard_normal((n_samples, n_features), chunks=(data_chunksizes, -1), dtype=dtype) else: # Randomly generate a low rank, fat tail input set if use_full_low_rank: X = make_low_rank_matrix(n_samples=n_samples, n_features=n_features, effective_rank=effective_rank, tail_strength=tail_strength, random_state=rs, n_parts=n_parts, n_samples_per_part=n_samples_per_part, dtype=dtype) X = X.rechunk({0: data_chunksizes, 1: -1}) else: seed = int(rs.randint(n_samples).compute()) covar = _make_low_rank_covariance(client, n_features, effective_rank, tail_strength, seed, n_parts, n_samples_per_part, dtype) X = _data_from_multivariate_normal(client, rs, covar, data_chunksizes, n_features, dtype) X = _convert_to_order(client, X, data_chunksizes, order, n_features, dtype) # Generate a ground truth model with only n_informative features being non # zeros (the other features are not correlated to y and should be ignored # by a sparsifying regularizers such as L1 or elastic net) ground_truth = 100.0 * rs.standard_normal((n_informative, n_targets), chunks=(n_samples_per_part, -1), dtype=dtype) y = da.dot(X[:, :n_informative], ground_truth) + bias if n_informative != n_features: zeroes = 0.0 * rs.standard_normal( (n_features - n_informative, n_targets), dtype=dtype) ground_truth = da.concatenate([ground_truth, zeroes], axis=0) ground_truth = ground_truth.rechunk(-1) # Add noise if noise > 0.0: y += rs.normal(scale=noise, size=y.shape, dtype=dtype) # Randomly permute samples and features if shuffle: features_indices = np.random.permutation(n_features) X, y = _shuffle(client, rs, X, y, data_chunksizes, n_features, features_indices, n_targets, dtype) ground_truth = ground_truth[features_indices, :] y = da.squeeze(y) if order == 'F' and n_targets > 1: y = _convert_to_order(client, y, y.chunks[0], order, n_targets, dtype) if coef: ground_truth = _convert_to_order(client, ground_truth, ground_truth.chunks[0], order, n_targets, dtype) if coef: ground_truth = da.squeeze(ground_truth) return X, y, ground_truth else: return X, y
def MDD(G, d, dt=0.004, dr=1., nfmax=None, wav=None, twosided=True, adjoint=False, dottest=False, saveGt=False, add_negative=True, **kwargs_cgls): r"""Multi-dimensional deconvolution. Solve multi-dimensional deconvolution problem using :py:func:`scipy.sparse.linalg.lsqr` iterative solver. Parameters ---------- G : :obj:`dask.array.ndarray` Multi-dimensional convolution kernel in frequency domain of size :math:`[n_{f,max} \times n_s \times n_r]` d : :obj:`dask.array.ndarray` Data in time domain :math:`[n_t \times n_s (\times n_vs)]` if ``twosided=False`` or ``twosided=True`` and ``add_negative=True`` (with only positive times) or size :math:`[2*n_t-1 \times n_s (\times n_vs)]` if ``twosided=True`` dt : :obj:`float`, optional Sampling of time integration axis dr : :obj:`float`, optional Sampling of receiver integration axis nfmax : :obj:`int`, optional Index of max frequency to include in deconvolution process wav : :obj:`numpy.ndarray`, optional Wavelet to convolve to the inverted model and psf (must be centered around its index in the middle of the array). If ``None``, the outputs of the inversion are returned directly. twosided : :obj:`bool`, optional MDC operator and data both negative and positive time (``True``) or only positive (``False``) add_negative : :obj:`bool`, optional Add negative side to MDC operator and data (``True``) or not (``False``)- operator and data are already provided with both positive and negative sides. To be used only with ``twosided=True``. adjoint : :obj:`bool`, optional Compute and return adjoint(s) dottest : :obj:`bool`, optional Apply dot-test saveGt : :obj:`bool`, optional Save ``G`` and ``G^H`` to speed up the computation of adjoint of :class:`pylops_distributed.signalprocessing.Fredholm1` (``True``) or create ``G^H`` on-the-fly (``False``) Note that ``saveGt=True`` will be faster but double the amount of required memory **kwargs_cgls Arbitrary keyword arguments for :py:func:`pylops_distributed.optimization.cg.cgls` solver Returns ------- minv : :obj:`dask.array.ndarray` Inverted model of size :math:`[n_t \times n_r (\times n_{vs})]` for ``twosided=False`` or :math:`[2*n_t-1 \times n_r (\times n_vs)]` for ``twosided=True`` madj : :obj:`dask.array.ndarray` Adjoint model of size :math:`[n_t \times n_r (\times n_{vs})]` for ``twosided=False`` or :math:`[2*n_t-1 \times n_r (\times n_r) ]` for ``twosided=True`` See Also -------- MDC : Multi-dimensional convolution Notes ----- Refer to :class:`pylops.waveeqprocessing.MDD` for implementation details. Note that this implementation is currently missing the ``wav`` and ``causality_precond=False`` options. """ nf, ns, nr = G.shape nt = d.shape[0] if len(d.shape) == 2: nv = 1 else: nv = d.shape[2] if twosided: if add_negative: nt2 = 2 * nt - 1 else: nt2 = nt nt = (nt2 + 1) // 2 nfmax_allowed = int(np.ceil((nt2 + 1) / 2)) else: nt2 = nt nfmax_allowed = nt # Fix nfmax to be at maximum equal to half of the size of fft samples if nfmax is None or nfmax > nfmax_allowed: nfmax = nfmax_allowed logging.warning('nfmax set equal to ceil[(nt+1)/2=%d]' % nfmax) # Add negative part to data and model if twosided and add_negative: if nv == 1: d = da.concatenate((da.zeros((nt - 1, ns)), d), axis=0) else: d = da.concatenate((da.zeros((nt - 1, ns, nv)), d), axis=0) d = d.rechunk(d.shape) # Define MDC linear operator MDCop = MDC(G, nt2, nv=nv, dt=dt, dr=dr, twosided=twosided, saveGt=saveGt) if dottest: Dottest(MDCop, nt2 * ns * nv, nt2 * nr * nv, verb=True) # Adjoint if adjoint: madj = MDCop.H * d.flatten() madj = da.squeeze(madj.reshape(nt2, nr, nv)) # Inverse minv = cgls(MDCop, d.flatten(), **kwargs_cgls)[0] minv = da.squeeze(minv.reshape(nt2, nr, nv)) #if wav is not None: # minv = sp_convolve1d(minv, wav, axis=-1) if adjoint: return minv, madj else: return minv
def segment_from_directory( directory, suffix, affinities_channels, centroids_channel, thresholding_channel, scale = (4, 1, 1), w_scale=None, compactness=0., display=True, validation=False, **kwargs # ): images, _, output, GT = get_dataset(directory, GT=True, validation=validation) images = da.squeeze(images) print(output.shape) segmentations = [] masks = [] scores = {'GT | Output' : [], 'Output | GT' : []} for i in range(output.shape[0]): gt = GT[i].compute() seg, _, mask = segment_output_image( output[i], affinities_channels, centroids_channel, thresholding_channel, scale=w_scale, compactness=0.) vi = variation_of_information(gt, seg) scores['GT | Output'].append(vi[0]) scores['Output | GT'].append(vi[1]) seg = da.from_array(seg) segmentations.append(seg) masks.append(mask) segmentations = da.stack(segmentations) masks = da.stack(masks) # Save the VI data scores = pd.DataFrame(scores) if validation: s = 'validation_VI.csv' else: s = '_VI.csv' s_path = os.path.join(directory, suffix + s) scores.to_csv(s_path) gt_o = scores['GT | Output'].mean() o_gt = scores['Output | GT'].mean() print(f'Conditional entropy H(GT|Output): {gt_o}') print(f'Conditional entropy H(Output|GT): {o_gt}') if display: # Now Display z_affs = output[:, affinities_channels[0], ...] y_affs = output[:, affinities_channels[1], ...] x_affs = output[:, affinities_channels[2], ...] c = output[:, thresholding_channel, ...] cl = output[:, centroids_channel, ...] v_scale = [1] * len(images.shape) v_scale[-3:] = scale print(images.shape, v_scale, z_affs.shape, masks.shape) v = napari.Viewer() v.add_image(images, name='Input images', blending='additive', visible=True, scale=v_scale) v.add_image(c, name='Thresholding channel', blending='additive', visible=False, scale=v_scale) v.add_image(cl, name='Centroids channel', blending='additive', visible=False, scale=v_scale) v.add_image(z_affs, name='z affinities', blending='additive', visible=False, scale=v_scale, colormap='bop purple') v.add_image(y_affs, name='y affinities', blending='additive', visible=False, scale=v_scale, colormap='bop orange') v.add_image(x_affs, name='x affinities', blending='additive', visible=False, scale=v_scale, colormap='bop blue') v.add_labels(masks, name='Masks', blending='additive', visible=False, scale=v_scale) v.add_labels(GT, name='Ground truth', blending='additive', visible=False, scale=v_scale) v.add_labels(segmentations, name='Segmentations', blending='additive', visible=True, scale=v_scale) napari.run()
def boundary_weights(self, mode='reflect', mask=None, drop_dims=[], compute=False): """ Compute the boundary weights Parameters ---------- mode : {'reflect', 'periodic', 'any-constant'}, optional The mode parameter determines how the array borders are handled. Default is 'reflect'. mask : array-like, optional Specify the mask, if None the mask is inferred from missing values drop_dims : list, optional Specify dimensions along which the weights do not need to be computed compute : bool, optional If True, the computation is performed after the dask graph has been made. If False, only the dask graph is made is the computation will be performed later on. Returns ------- weights : xarray.DataArray Return a DataArray containing the weights """ if mode is 'periodic': mode_conv = 'wrap' else: mode_conv = mode # Normalize coefficients coeffs = self.coefficients / self.coefficients.sum() if drop_dims: new_coeffs = da.squeeze( coeffs, axis=[self.obj.get_axis_num(di) for di in drop_dims]) else: new_coeffs = coeffs new_obj = self.obj.isel(**{di: 0 for di in drop_dims}).squeeze() #depth = {new_obj.get_axis_num(di): self.order[di] // 2 # for di in self.dims} boundary = {self._obj.get_axis_num(di): mode for di in self.dims} if mask is None: mask = da.notnull(new_obj.data) conv = lambda x: im.convolve(x, new_coeffs, mode=mode_conv) weights = mask.astype(float).map_overlap(conv, depth=self._depth, boundary=boundary, trim=True) res = xr.DataArray(mask * weights, dims=new_obj.dims, coords=new_obj.coords, name='boundary_weights') res = res.where(res != 0) if compute: with ProgressBar(): out = res.compute() else: out = res return out
def main(argv=None): # cluster = LocalCluster(dashboard_address=None) # client = Client(cluster, memory_limit='{}GB'.format(FLAGS.memory_limit), # processes=False) K.set_floatx('float32') chunk_size = FLAGS.chunk_size # Read data set hdf5_file = h5py.File(FLAGS.data_file, 'r') images, labels, _ = hdf52dask(hdf5_file, FLAGS.group, chunk_size, shuffle=FLAGS.shuffle, seed=FLAGS.seed, pct=FLAGS.pct) n_images = images.shape[0] n_batches = int(np.ceil(n_images / float(FLAGS.batch_size))) # Data augmentation parameters daug_params_file = get_daug_scheme_path(FLAGS.daug_params, FLAGS.data_file) daug_params = yaml.load(open(daug_params_file, 'r'), Loader=yaml.FullLoader) nodaug_params_file = get_daug_scheme_path('nodaug.yml', FLAGS.data_file) nodaug_params = yaml.load(open(nodaug_params_file, 'r'), Loader=yaml.FullLoader) # Initialize the network model model_filename = FLAGS.model model = load_model(model_filename) # Print the model summary model.summary() # Get relevant layers if FLAGS.store_input: layer_regex = '({}|.*input.*)'.format(FLAGS.layer_regex) else: layer_regex = FLAGS.layer_regex layers = [ layer.name for layer in model.layers if re.compile(layer_regex).match(layer.name) ] # Create batch generators n_daug_rep = FLAGS.n_daug_rep n_diff_per_batch = int(FLAGS.batch_size / n_daug_rep) image_gen_daug = get_generator(images, **daug_params) batch_gen_daug = batch_generator(image_gen_daug, images, labels, batch_size=n_diff_per_batch, aug_per_im=n_daug_rep, shuffle=False) image_gen_nodaug = get_generator(images, **nodaug_params) batch_gen_nodaug = batch_generator(image_gen_nodaug, images, labels, FLAGS.batch_size, aug_per_im=1, shuffle=False) # Outputs if FLAGS.output_dir == '-1': FLAGS.output_dir = os.path.dirname(FLAGS.model) output_hdf5 = h5py.File( os.path.join(FLAGS.output_dir, FLAGS.output_mse_matrix_hdf5), 'w') output_pickle = os.path.join(FLAGS.output_dir, FLAGS.output_pickle) df_init_idx = 0 df = pd.DataFrame() # Iterate over the layers for layer_idx, layer_name in enumerate(layers): # Reload the model if layer_idx > 0: K.clear_session() model = load_model(model_filename) layer = model.get_layer(layer_name) # Rename input layer if re.compile('.*input.*').match(layer_name): layer_name = 'input' hdf5_layer = output_hdf5.create_group(layer_name) activation_function = K.function( [model.input, K.learning_phase()], [layer.output]) print('\nComputing pairwise similarity at layer {}'.format(layer_name)) # Compute activations of original data (without augmentation) a_nodaug_da = get_activations(activation_function, batch_gen_nodaug) a_nodaug_da = da.squeeze(a_nodaug_da) a_nodaug_da = da.rechunk(a_nodaug_da, (chunk_size, ) + (a_nodaug_da.shape[1:])) dim_activations = a_nodaug_da.shape[1] # Comute matrix of similarities r = da.reshape(da.sum(da.square(a_nodaug_da), axis=1), (-1, 1)) mse_matrix = (r - 2 * da.dot(a_nodaug_da, da.transpose(a_nodaug_da)) \ + da.transpose(r)) / dim_activations # Compute activations with augmentation a_daug_da = get_activations(activation_function, batch_gen_daug) a_daug_da = da.rechunk(a_daug_da, (chunk_size, dim_activations, 1)) # Compute similarity of augmentations with respect to the # activations of the original data a_nodaug_da = da.repeat(da.reshape(a_nodaug_da, a_nodaug_da.shape + (1, )), repeats=n_daug_rep, axis=2) a_nodaug_da = da.rechunk(a_nodaug_da, (chunk_size, dim_activations, 1)) mse_daug = da.mean(da.square(a_nodaug_da - a_daug_da), axis=1) # Compute invariance score mse_sum = da.repeat(da.reshape(da.sum(mse_matrix, axis=1), (n_images, 1)), repeats=n_daug_rep, axis=1) mse_sum = da.rechunk(mse_sum, (chunk_size, 1)) invariance = 1 - n_images * da.divide(mse_daug, mse_sum) print('Dimensionality activations: {}x{}x{}'.format( n_images, dim_activations, n_daug_rep)) # Store HDF5 file if FLAGS.output_mse_matrix_hdf5: mse_matrix_ds = hdf5_layer.create_dataset( 'mse_matrix', shape=mse_matrix.shape, chunks=mse_matrix.chunksize, dtype=K.floatx()) mse_daug_ds = hdf5_layer.create_dataset('mse_daug', shape=mse_daug.shape, chunks=mse_daug.chunksize, dtype=K.floatx()) invariance_ds = hdf5_layer.create_dataset( 'invariance', shape=invariance.shape, chunks=invariance.chunksize, dtype=K.floatx()) time_init = time() with ProgressBar(dt=1): da.store([mse_matrix, mse_daug, invariance], [mse_matrix_ds, mse_daug_ds, invariance_ds]) time_end = time() print('Elapsed time: {}'.format(time_end - time_init)) invariance = np.ravel( np.asarray(output_hdf5[layer_name]['invariance'])) else: time_init = time() invariance = da.ravel(invariance).compute() time_end = time() print('Elapsed time: {}'.format(time_end - time_init)) # Update pandas data frame for plotting df_end_idx = df_init_idx + n_images * n_daug_rep d = pd.DataFrame( { 'Layer': layer_name, 'sample': np.repeat(np.arange(n_images), n_daug_rep), 'n_daug': np.tile(np.arange(n_daug_rep), n_images), 'invariance': invariance }, index=np.arange(df_init_idx, df_end_idx).tolist()) df = df.append(d) df_init_idx += df_end_idx pickle.dump(df, open(output_pickle, 'wb')) output_hdf5.close()